Nitin Kanukolanu

Agents don’t flow — they live. I keep seeing new “agent” frameworks that simply repackage workflows — connecting steps in a DAG (Directed Acyclic Graph) and calling it autonomy. The latest example: OpenAI’s AgentKit announced at Dev Day → https://lnkd.in/g7FCW6Mt But autonomy isn’t something you add to a workflow — it’s something that defines an agent. True agents don’t flow from one step to the next. They exist in environments — perceiving, deciding, and acting continuously based on what they know, sense, and aim to achieve. When we treat agents as workflow nodes, we lose what makes them agentic in the first place: persistence, specialization, and emergent coordination. Sustainable architectures won’t come from orchestrated flows dressed up as agents. They’ll come from systems where intelligence arises from interaction, not instruction. (Not a critique of any vendor — just a reminder that autonomy isn’t a feature you wire in; it’s the nature of an agent itself.) #AI #Agents #SystemsArchitecture #Autonomy #MultiAgentSystems #LLM #EmergentIntelligence #AgenticAI #Innovation #DesignPhilosophy

60

Great time at the Liquidity Summit two weeks ago. I had the chance to sit down with the NYSE right after Bill Ackman to talk AI investment trends, compute demand, and what's driving the next phase of growth. Here's the conversation.

93

1 Comment

What if I “wrote” the entire app in the Markdown instructions file? Great post on spec-driven design using Github’s spec kit — https://lnkd.in/giBinW7f As ever, another great post from Anthropic on the use of code to drive MCP reasoning logic and thereby greatly improve token-use efficiency [present MCP servers as code APIs rather than direct tool calls] — https://lnkd.in/gRmYjG6i And, talking of MCP token efficiency, this open source multi-agent example from Tiger Data (creators of TimescaleDB) is a great resource — https://lnkd.in/gBkZSEc2

12

If you are still using plain JSON for your LLMs, you might be missing out on significant savings in tokens, speed, and budget. TOON, launched last week, is a transformative serialization format designed for the era of large context, agentic workflows, and multi-tool LLM systems. Here’s what TOON offers: - Token efficiency: Reduces token usage by 30-60% compared to formatted JSON for uniform arrays. - LLM-friendly structure: Features explicit lengths and field sets that enhance parsing and retrieval accuracy. - Minimal syntax: Eliminates excess braces, quotes, and repeated keys, resulting in cleaner input for quicker comprehension. - Seamless JSON compatibility: Allows you to spend in JSON, push to TOON for LLMs, and convert back when necessary. When to use TOON: - Best suited for large datasets with uniform object arrays, such as telemetry logs, user lists, and product catalogs. - Less effective for deeply nested, non-uniform structures, where JSON may still be preferable. For those developing agents, copilots, or any LLM system that sends structured data into a prompt, adopting TOON as your standard is highly recommended. Your token budget, model throughput, and reliability will benefit significantly.

15

Cassandra (Cassie) Shum, VP of Ecosystem Product Engineering at RelationalAI, on whether LLMs have made knowledge graphs redundant: The retrieval argument for graphs was stronger six months ago. Models have caught up on that dimension. The more durable argument is treating the knowledge graph as the foundation of your system rather than as a query layer. More sessions like this are lined up for QCon AI New York 2026 (Dec 15-16). Find out more about QCon AI New York here https://bit.ly/4yDXymz

1

AI Doesn’t Transform Companies. Decision Rights Do. Post 1 — Map the Decisions Before You Map the Models Most AI programs start with a roadmap of systems. Very few start with a roadmap of decisions. We obsess over: • which LLM • which vendor • which pilots But rarely over: • which decisions will this AI actually own? • which ones will it only recommend on? • which ones are too high-risk to touch (for now)? That’s why so many “AI transformations” quietly degrade into dashboards, copilots, and sidecar tools: impressive tech, unchanged decision rights. Every serious AI program needs a decision map Before you scale models, you need to see the decision landscape of your organization. I like to start with three simple questions: 1. What are the highest-volume decisions we make every day? (approvals, routing, triage, pricing tweaks, risk checks, etc.) 2. Who actually makes them today? (not the org chart — the real deciders: analysts, ops leads, “that one person” everyone Slacks) 3. What’s the risk if this decision goes wrong? (inconvenience, lost revenue, regulatory impact, reputational damage) Once you have this, you can start to classify: • Low-risk → candidates for AI ownership • Medium-risk → candidates for AI within guardrails • High-risk → candidates for AI assistance + escalation Now you’re not “doing AI.” You’re redesigning how the company decides. A quick example Take a simple one: refund decisions in a digital business. • Today: - frontline teams follow a fuzzy playbook - “exceptions” go to managers - customers wait - decisions are inconsistent • With a decision map, you might define: - Low-risk: small refunds under a certain threshold → AI can auto-approve using clear rules and patterns - Medium-risk: larger amounts → AI proposes a decision within policy guardrails, manager reviews with full context - High-risk: suspicious patterns, potential fraud, VIP customers → AI surfaces context, human owns the call Same AI capability. Completely different impact when decision rights are explicit. The uncomfortable truth If you skip this step, your AI will quietly become: “A very smart advisor… waiting on the same old approvals.” Post 2 in this series will go deeper into tiered autonomy — own, guardrail, escalate — and how to apply it without losing sleep (or control). Until then, a question for you: If you listed your top 10 daily decisions, how many of them have clear owners and risk levels today? #DecisionIntelligence #AILeadership #AutonomousEnterprise #AIGovernance #OperatingModel #CIO #CTO

23

Here's how I explain LLMs to non-technical clients. LLMs change traditional software by enabling 3 things... 1) Understanding messy, variable input 2) Producing outputs that aren’t predefined and adapt to the task & context 3) Generalizing across edge cases that used to require rigid rules or cleanup Basically, flexibility: the part humans would usually handle. Most explanations overcomplicate this. How’s my explanation? How do you explain LLMs non-technically?

9

1 Comment